Severe convective weather, such as short-time heavy precipitation and strong wind, has great harm, and there are considerable technical difficulties in its automatic identification. An intelligent recognition model of severe convective weather based on deep neural network is proposed, the model takes radar echo image and optical flow image representing echo movement path as input. Through self-learning of neural network, the functional mapping relationship between radar image and whether severe convection weather occurs is sought. The techniques of data set enhancement, cost function optimization and model generalization performance optimization are used to solve the problem of unbalanced training samples and avoid the problem of model training falling into local extremum. Experimental results show that the accuracy of this method is 96%, and the false alarm rate is less than 60%. This method is also suitable for automatic identification of severe weather such as downburst.